This study aims to develop a novel approach of combining machine learning with physics-based mathematical models for near-real-time clear-sky irradiance prediction using high temporal resolution geostationary satellite (GOES-16) imagery. The proposed method produces a spatiotemporal data stream of 2 km resolution at 5-minute granularity, allowing the mathematical models (REST2V5, MAC2) to estimate clear-sky irradiance based on the unique nuances of the local meteorological characteristics. Artificial neural networks and gradient-boosting models are then trained using the mathematical model outputs as labels. Utilizing the least number of meteorological features based on their importance, our approach predicted clear-sky irradiance for three different sunny days, achieving the coefficient of determination as 0.94 with ground-based weather station (SURFRAD) data, which was around 10% higher compared to the prediction from global reanalysis datasets (MERRA2, CAMS). The proposed approach also reduced the computational complexity by decreasing the dimensionality of the required data.

Floods pose significant threats to communities, economies, and ecosystems worldwide. Effective flood monitoring and management systems are crucial for mitigating the devastating impacts of these natural disasters. Although satellite observation, particularly using Synthetic Aperture Radar (SAR) sensors, has greatly enhanced flood monitoring capabilities, near-real-time monitoring still faces challenges in automatically detecting floods. This study aims to develop an efficient approach for near-real-time flood detection from satellite observation data without relying on training a model. We utilize a Bayesian analysis of Change Point problems (BCP) to obtain a probability of change at a pixel-level from newly observed data. We demonstrate this approach using openly available Sentinel-1 time series data and assess its effectiveness using the Sen1Flood11 dataset. Our method successfully detects floods in the latest observation, distinguishing them from permanent water areas. Future work includes addressing false positives and false negatives in flood detection and exploring the application of our approach to a wider range of flood events.

We performed flood inundation mapping using polarimetry decompositions on Sentinel-1 synthetic aperture radar data for the 2019 Hagibis flood event in Ibaraki, Japan. We have proposed an index-based method based on Stokes parameters obtained from dual-polarized Sentinel-1 data. Our method shows that the eigenvalue λ1-based flood inundation (FI) map achieves an F1 score of approximately 0.73 compared to a high-resolution flooded optical image.

This study presents a method for mapping forest vertical structures using fused satellite data sets, including data from Global Ecosystem Dynamics Investigation (GEDI) mission, Sentinel-1 and -2, by a Random Forest classifier. The method aims to develop an effective yet simple way to map the foliage height diversity, plant area index, canopy height, and forest structure index, which is a composite index of these three metrics. They are mapped at 10 m spatial resolution, which can provide information about the distribution and functioning of different plant functional types and canopy layers in a forest. The approach was tested in an area around Mt. Washibetsu, Hokkaido, Japan, and demonstrated the feasibility of capturing forest vertical structures using satellite remote sensing, which has important implications for forest management and conservation.

The objective of this study was to develop a 10-meter resolution land cover classification map using Sentinel-1 and Sentinel-2 satellite imagery along with the Dynamic World data set. A Gradient Boosted Decision Tree (GBDT) classification was employed with reference samples. To reduce the salt-and-pepper effect, a decision fusion approach was applied to the classification probabilities obtained from the GBDT and the CNN-based classification from Dynamic World. The model’s accuracy increased when CNN-based classification probabilities were used as inputs of the GBDT classification, while the salt-and-pepper effect was reduced when a decision fusion was applied to those two probabilities. This study suggests that the traditional land cover classification approach with reference sampling can still be effective when integrating with CNN-based classification probabilities for rapid land cover classification mapping.

Although geographically weighted Poisson regression (GWPR) is a popular regression for spatially indexed count data, its development is relatively limited compared to that found for linear geographically weighted regression (GWR), where many extensions (e.g. multiscale GWR, scalable GWR) have been proposed. The weak development of GWPR can be attributed to the computational cost and identification problem in the underpinning Poisson regression model. This study proposes linearized GWPR (L-GWPR) by introducing a log-linear approximation into the GWPR model to overcome these bottlenecks. Because the L-GWPR model is identical to the Gaussian GWR model, it is free from the identification problem, easily implemented, computationally efficient, and offers similar potential for extension. Specifically, L-GWPR does not require a double-loop algorithm, which makes GWPR slow for large samples. Furthermore, we extended L-GWPR by introducing ridge regularization to enhance its stability (regularized L-GWPR). The results of the Monte Carlo experiments confirmed that regularized L-GWPR estimates local coefficients accurately and computationally efficiently. Finally, we compared GWPR and regularized L-GWPR through a crime analysis in Tokyo.

Although the scalable geographically weighted regression (GWR) has been developed as a fast regression approach modeling non-stationarity, its potential on spatial prediction is largely unexplored. Given that, this study applies the scalable GWR technique for large-scale spatial prediction, and compares its prediction accuracy with modern geostatistical methods including the nearest-neighbor Gaussian process, and machine learning algorithms including light gradient boosting machine. The result suggests accuracy of our scalable GWR-based prediction.

Floral phenology is useful information as an indicator of climate change and ecosystem services, however its observation is not straightforward over space and time. Satellite remote sensing and official and volunteer-based in-situ observations have been conducted, but the long-term and accurate data collection is challenging due to the insufficient quality and quantity of observations and the lack of financial and human resources to sustain them. Here, we demonstrate a flower detection model from street-level photos, which can be the core function of a semi-automatic observation system to tackle those issues above. We detected cherry blossoms by this model from geotagged images with the observation date, obtained from Mapillary, which is one of the social sensing data sources, and mapped dates of flowering in a study site, Aizuwakamatsu, Japan in April 2018. This approach enables us to collect floral phenology information semi-automatically as a data-driven approach. It is expected to collect a large number of observations with a certain level of quality by avoiding human-induced biases in the observations.

Land cover classification mapping is a process to depict terrestrial surfaces by pre-defined thematic classes across space and is often implemented by a supervised classification approach with satellite and/or aero images. However, due to labor-intensive and time-consuming tasks to identify land cover class from those images manually, it is challenging to build richquantity and highquality reference data sets for the mapping. To capture land covers on the ground, we developed a deep learning model to estimate land covers from street-level photos. A transfer learning from the pre-trained DenseNet was applied, and this model yields an overall accuracy of 0.91. This model contributes to the subsequent study of building a semi-automatic reference database from geo-tagged street-level photos.

Errors in land cover classification are often spatially heterogeneous even though a soft classification model such as spectral unmixing is implemented to mitigate a mixed pixel problem. The estimated land covers are fractions of targeted classes with the restriction of the sum to one and being non-negative. To assess the classification with considering a spatial heterogeneity, we propose a geographically weighted total composite error analysis. By using the USGS global reference database, we assessed errors of spectral unmixing classification of ALOS AVNIR-2 data into 4 land cover classes. Results yield a spatial surface of local errors by the Aitchison distance and address that the error magnitude across space is associated with the complexity of land covers.